Man’s exploitation of the earth’s hydrocarbons brought extraordinary benefits, but also led to the First Great Pollution. Belatedly we have learned that burning hydrocarbons has polluted and changed our climate so as to become a danger to all life, and plastics from hydrocarbons are endangering oceanic life.
We have now agreed that we must stop this pollution, but agreeing to stop is the easy bit. Actually stopping, let alone reversing the heating and the pollution will be at best phenomenally expensive, at worst impossible.
The Second Great Pollution, of all the data and information that is stored in our technology and transmitted over the internet (what we call the ‘infosphere’), only started in this century. Although we humans originate this pollution, the scale of harm it causes is entirely due to its spreading and amplification by social media and artificial intelligence (AI).
Social media have brought undoubted benefits, and the enthusiasts for AI promise us a brilliant future, but the dangers and harms caused by these technologies are now becoming abundantly clear, even though they are less tangible than the harms from exploiting hydrocarbons. We must act now, drawing on lessons from the First Great Pollution. before it is too late to prevent another disaster.
1. The way information is spread has changed in the 21st Century
Before the advent of social media and AI, all public discourse, whether on facts, news, opinions, rumours or whatever, took place face-to-face or via print media (books, newspapers, journals, etc.), or radio and TV. Use of the internet for e-mail and search had started but was not yet widespread.
This discourse included mis- and dis-information, insults and lies (and of course always will) but there were several important natural constraints that limited the harm they could cause.
Most importantly, 20th Century communication was inherently one-to-many. This meant that information receivers generally knew the sources and had some feeling for how much they could be trusted. Further, these information sources were, and are still, subject to a variety of constraints. Public broadcasters must abide by licensing conditions; libel and slander laws inhibit private media owners’ from taking too much risk; cost is a significant barrier for anyone to enter this space.
However over the last 10 – 20 years, social media and AI have changed the scale and nature of public discourse. Now, every individual who can afford a smartphone can add to the infosphere on any topic, broadcasting their views to everyone anywhere. Communication can now be many-to-many, with the original source usually unknown to the recipients, or deliberately hidden. While much of this social media traffic is entirely innocent, an increasing proportion can only be described as ‘infospheric pollution’, as illustrated below.
Beyond the widely-reported cases of LLMs giving advice based on biased data, and of hallucinating, there are other dangers to users of LLMs and of AI-assisted search tools. Their astonishing capability to rapidly produce coherent, credible answers to any question means we are more likely to accept their answers than to critically examine them for correctness and completeness.
Some examples of AI untrustworthiness from the author’s own experience:
- Superficial answers. In August 2023, I planned to present a paper at a conference in Rome in early October. My UK passport expired at end February 2024, so I checked on the rules for its validity for entry to the EU. MS Edge told me it must have ‘3 unexpired months’. OK, so I booked and paid for my flights and hotel. At the departure gate I was told I could not fly. Edge had failed to mention another obscure rule which meant my passport was not valid for entry to the EU. My loss ~ US$750. Charles Symons
- Varying answers. As an experiment, two friends A and B, searched Google for news about a recent election rumour, using identical wording, at about the same time. They got strikingly different answers. The answer to friend A who was in Oklahoma would have pleased a Republican voter; the answer to the friend B who was in New York would have pleased a Democratic voter. What/who can we trust nowadays? Ben Porter.
- Generated code As IT professionals with experience dating back decades, we are very concerned about AI-generated program code. We fear that the ease with which code can be generated, combined with the pressures of DevOps processes, will result in code receiving only cursory human examination being added to production systems – with who-knows-what short- or long-term consequences? Charles and Ben
2. Social media and AI suppliers’ profit goals clash with their efforts to prevent harmful traffic
Efforts by social media suppliers to curb harmful pollution on their systems are at best inadequate. At worst, their AI-assisted algorithms amplify the harms by encouraging (sometimes even rewarding) content that leads to greater engagement, and eventually to addiction. The more controversial or toxic a posting, the greater the engagement, attracting more revenue from advertisements, and so more profit for the supplier, regardless of any resulting harms.
As a further means of encouraging engagement, individual recipients’ personal interests in particular topics or viewpoints are recognized by the social media AI algorithms so that each recipient is fed more material in line with their interests. It thus becomes impossible to get a balanced view of what’s going on in the world from these sources, while at the same time existing biases and prejudices are automatically re-enforced.recipients’ personal interests in particular topics or viewpoints are recognized
Large Language Models (LLMs) and AI search tools are now polluting the infosphere in other ways. Their output of biased data, superficial, variable, or wrong answers and hallucinations, may be sucked up into the tools’ next learning cycles, unfiltered by any concept of accuracy or truth, with ever-more harmful consequences.
3. The greatest harm to society from this Second Great Pollution is the undermining of trust.
Trust in information about individuals and institutions is the foundation of a healthy society. Breakdown of trust by the spread and amplification of mis- and dis-information (including conspiracy theories), lies and vitriol increases tensions and divisions in society, ultimately posing a threat to democracy.
Big Brother (from George Orwell, ‘1984’) would have loved to have owned ‘X’!
This problem becomes ever more acute as more and more people turn to AI-manipulated social media and search engines for their news and information rather than traditional media curated by professional journalists. As individuals are increasingly exposed to a polluted infosphere, they may be left angry or confused and anxious, or to their tuning out to avoid the pollution, leaving them ignorant. This is leading to rising mental health problems, particularly for the young.
4. Lessons from the way society is tackling the First Great Pollution.
These mechanisms for spreading harm cannot be allowed to continue unchecked. First, we need to accept that the harms caused by social media and AI technology are not inevitable. If that means regulations, with punishments for those who won’t curb their technology, then so be it.
Opponents of regulation cite their rights to freedom of expression and that social media simply provide the modern equivalents of Town Hall meeting places. This view is just not true.
Unlike the 18th Century when this principle was first defined, social media sites are not forums for open discussion where anyone can publicly express an opinion which anyone else can hear and challenge, face-to-face.
As we have seen, discourse in the infosphere is controlled by the social media algorithms whose main goal is to maximise profit for the supplier. Much of the input to this discourse is posted anonymously and much is hidden behind encryption. Further, freedom of expression has never been regarded as an absolute right in all circumstances. Most countries have laws protecting people from slander, libel or invasion of privacy; the USA has laws prohibiting calls for insurrection or for causing unnecessary alarm to the public. So why should social media be a protected forum for complete freedom of expression?
Just as the right to bear arms does not extend to the right to shoot anyone you dislike, so the right to freedom of expression should not extend to anyone to use words or images to seriously harm others, combined with technology to spread their poison to the entire world from behind a wall of anonymity.
Tackling these harms must put the interests of protecting individuals and society above all other considerations. This will require strong leadership to overcome opposition to regulations. Happily, the EU, the UK and the USA (thus far) are taking the first steps towards agreeing the need for regulations, under the heading of ‘AI Safety’.
However, it is already clear that it is extremely difficult to identify and remove all harmful content posted on social media, even using AI supported by human moderators. So in addition to regulations, we need the equivalents of carbon capture to attenuate or remove the pollution and filters that enable users to protect themselves from its harmful effects. We will also need the equivalent of green energy to replace the polluting sources.
5. Possible steps to counter infospheric pollution or to mitigate its effects
a. Stem the pollution spread by social media
In the USA, a great start would be to remove the protection afforded by Section 230 of the Communications Decency Act that absolves social media from any liability for the traffic they carry. This Act was introduced in 1996 to give all internet users the same protection as telephone carriers. This traffic had historically comprised one-to-one voice calls, over which carriers had no control. This reasoning is no longer justifiable when social media systems carry and moderate their traffic, their algorithms determining which postings are amplified and which get no air.
Removing this protection would open the social media companies to liability for compensation to their users who suffer fraud or other harms, providing a big incentive for the companies to curb pollution. A specific anti-pollution step aimed at protecting young persons is the proposed ‘Kids On-line Safety Act’. At the time of writing, this bill is currently in Congress, held up by lobbying from social media companies aimed at limiting their duty of care.
In other countries, e.g. the EU and the UK, legislation is being introduced to require social media companies to implement stronger content moderation under threat of very serious fines for non-compliance.
Other measures worth considering would be aimed at reducing the amount of heat generated by social media and thus their influence. Examples include providing for ‘Dislikes’, ‘Untrue’, or such-like (in addition to ‘Likes’), and measures for attenuating the spread of messages, e.g. by limiting each user to a small number of posts and re-posts per day.
b. Stem the pollution generated by LLM’s
The principal effort needed to stem this form of pollution must come from the LLM suppliers’ testing of their own products, to prevent them from generating misinformation or hallucinations. One simple step would be to prohibit artificial data generated by LLM’s for testing purposes from being made public in the infosphere.
All LLM output should be required to be indelibly ‘watermarked’, should provide (on demand) references for its factual replies, and to show its reasoning for derived replies.
Another step with more subtle effects would be to stop LLMs ‘breathing their own or other LLMs’ exhaust’, i.e. incorporating their own output, in each of their successive learning cycles. This should have two beneficial effects. First, it would prevent earlier learning that was correct and/or historically important from being down-graded or over-written by later cycles. Second, it should reduce ‘regression to the mean’ in LLM output. Tests have shown that repeated cycling results in the production of text that is at best predictable and dull, at worst gibberish, and that a selection of images, e.g. of different faces, will be reduced to one common synthesized image.1
c. Enable users to protect themselves
Social media systems should be required to provide two main mechanisms to allow users to protect themselves from harmful material.
First a standard API should be defined and made available by all social media systems to allow third-party software to be installed on the user’s own device so that users can shield themselves from harmful input when scrolling or in a dialogue2. A market for the provision of such shielding software would surely grow rapidly.
Sentiment analysis has the potential to label each post with a ‘heat index’ enabling a user to filter out e.g. extreme views. However, social media suppliers are unlikely to implement any feature that is against their commercial interests. And the compute power needed to apply this approach in real time to all traffic to a user’s device probably rules it out for the time being. Maybe a measure for the longer-term?
Second, specific measures are needed to protect young people from the harms of social media. In particular, social media must incorporate strong age-verification functionality, enabling them to restrict material that is unsuitable for younger users.
More generally, educating young people about the dangers of relying on information received via social media and from AI, and of addiction, is more important than ever. Banning the use of smartphones in schools would bring additional benefits, helping students to concentrate more on their studies, to spend more time interacting socially with their class-mates, and to reduce their addiction to social media.
Australia has taken the lead in this action; other countries and individual schools are following this lead.
d. Clean up existing pollution
Perhaps this is where AI could yet render its greatest service to mankind by detecting and eliminating mis- and dis-information, fakes, lies, etc?
e. The ultimate sanction for persistent polluters
If power stations that burn fossil fuels can be phased out, then it should not be difficult to ban social media systems that continue to pollute the infosphere, harming society and individuals.
6. Longer-term risks of AGI
Beyond today’s concerns, Geoffrey Hinton, who shared a 2024 Nobel Prize for his work on laying the foundations for LLMs, warns of the longer-term risks of Artificial General Intelligence (AGI). Existing AI can already make better decisions than humans in specialist fields such as detecting cancers, and LLMs already ‘know’ vastly more than any individual. So Hinton suggests that the day may arrive when AGI is so powerful that it is asked to advise on policy decisions.
But how could this advice be trusted when the AGI is based on a polluted knowledgebase?
7. The final irony: the Second Great Pollution worsens the First Great Pollution!
Who would have expected pollution of the intangible infosphere to exacerbate pollution of the tangible physical environment?
Data centres currently account for about 3.5% of US electricity consumption. This share is forecast to rise to 9% by 20303 due to the demands of AI. Even if some of this increase in power comes from low-carbon generators, the trend is madness when we should be reducing our energy consumption, and the evidence is that LLMs are reaching their limits.
Acknowledgements
We are grateful to the members of the Technology, Media and Society group of Prometheus Endeavor for their input to this paper.
[1] When A.I.’s Output Is a Threat to A.I. Itself – The New York Times ,25th August 2024
[2] BlueSky has introduced such a filter using open-source software.
[3] Data and forecast from Barclays Investment Bank, August 28, 2024. The forecasters claim to have allowed for assumed efficiency gains of AI chips.
Excellent article and insights.
A very timely article. The irony you mention in closing – that AI will create massive data centers that will contribute to atmospheric pollution – is very discouraging. We seem to be unable to deal with our own self-destructive tendencies.